尺度无关的级联卷积神经网络人脸检测算法

Since the convolution neural network needs to input a fixed size image when performing image processing, this will lead to the loss of most of the original information in the scaling process. In addition, the feature extraction of images will not be put in place when the network has only one structure. To solve the above two problems, this paper presented a face detection algorithm combining cascade convolution neural network and spatial pyramid pooling. In this method, it cascaded three different convolution neural network models, the structure of which were from simple to complex, and extracted different face features at different levels to complete the detection of the face areas of images. At the same time, it added the pyramid pool at each level of the network, and this pooling strategy did not require a fixed size input, increasing dimension selection of model input. Compared with the traditional method, this method realizes the multi-scale input of the model, improves the detection performance, and reduces the time of face detection in the standard face data set.